Migration and Economic Recovery During COVID-19¶

Executive Summary¶

My project will analyze migration patterns and economic recovery following the COVID-19 pandemic to understand the potential relationship between the two and attempt to predict future migration trends.

The ability to predict which areas of the US will grow in population faster than others after a national crisis is especially relevant to policymakers, who must prioritize government aid to the areas that need it most. During COVID, there were billions of dollars of federal money given to states and cities to help them recover from the pandemic, leaving local policymakers to make decisions about where to best allocate this spending.[1] While COVID-19 may seem like a rare scenario, a recent study by the Proceedings of the National Academy of Sciences warns that similar pandemics will become more common in the future.[2] If policymakers are able to predict migration flows at a county-level, they will have a better sense of how to allocate aid on a per capita basis that reflects where people will be located in the future. Aid may also be more useful in areas that will continue to lose a lot of their population, which will inevitably affect the local economy. Furthermore, if these trends hold in post-COVID settings, predicting regional population growth could prove useful for other, non-crisis government initiatives, like infrastructure projects.

To accomplish these tasks, I employ county-level economic data that measure the impact of the COVID-19 pandemic on consumer spending, employment, and small businesses. Additionally, I use data on post office change of address requests to estimate migration patterns at the county-level. The result of this project is a model that can predict migration patterns, primarily through seasonal trends, but with slightly increased accuracy using economic data. I find that there are strong correlations between economic activity and migration patterns, though these relationships only lead to modest improvements in a model that attempts to predict future migration patterns.

Problem Statement¶

The COVID-19 pandemic caused an unprecedented economic shutdown across the United States, weakening the economy and closing businesses for months.[3] Simultaneously, the rise of remote work enabled workers to move further away from their office location without penalty. Previous work on this subject has found that during the pandemic, people moved away from central cities and into surrounding suburbs.[4] This raises the question of whether areas with greater migration inflow are also places that experienced less economic damage during 2020, as people were possibly drawn to areas due to the lower COVID impact on the local economy. If so, we may be able to use this same data to predict migration trends in the short-run, especially now that many workers are no longer confined to living near their company headquarters and can more easily relocate to more prosperous areas.

This project paints a picture of migration and economic recovery during COVID across the United States and uses this data to predict migration trends post-COVID. In the modern day, we have large amounts of data on economic activity that can be accessed quickly and used during crises to inform policy decisions. Knowing which variables are most useful in predicting future migration trends will be essential in responding quickly to these crises. Additionally, if we can leverage economic and migration data to predict future trends, this can be another tool for policymakers when designing more standard economic policies.

For this project, I use data from two sources. First, I gather data from Opportunity Insights.[5] Opportunity Insights is a group of researchers that uses economic data to explore inequalities across the United States. During COVID-19, this group partnered with several companies to track how the pandemic had affected local economies in the US and subsequently track the local recovery as the pandemic progressed. Specifically, they have data on county-level credit/debit card spending, employment, small business revenues and openings, and cellphone mobility. Second, I gather migration data from the United States Postal Service (USPS).[6] USPS collects data on monthly address changes at the ZIP code level. For each month, they report the total address changes into a ZIP code and the total address changes out of a ZIP code, allowing me to approxmiate how many people are moving to and away from each ZIP code.

The speed at which this data can be collected and processed is extremely powerful. For example, Opportunity Insights posts data on daily credit/debit card spending at the county-level within a few weeks of the activity. Additionally, monthly USPS data becomes available soon after the month ends, meaning we can use very recent data on economic activity and migration to predict future trends. One of the goals of this project, therefore, is to explore the potential of this data to make predictions in the near future, which will hopefully make those predictions more accurate.

Using these data sources and with the above motivation, this analysis sets out to answer the following questions:

1. What is the correlation between population movement, as tracked through USPS, and economic indicators? Did areas that saw increased migration post-COVID also experience less economic damage?
2. Which variables in these data are the strongest predictors of migration?
3. Can we use a combination of migration and economic data to project migration patterns in the short-term?
4. Which model is best for predicting migration patterns at the county-level?

To achieve these results, I use a combination of exploratory analysis, hypothesis testing, ordinary least squares regressions, L1/L2 regularization, and STL decomposition techniques.

Based on the initial exploratory data analysis, there appear to be correlations between migration to an area and that area's economic recovery, though there is substantial variation by state. This is expected, given differences between states in lockdown severity and length during the onset of COVID-19. The relationships between economic variables and migration also appear to vary over time in some cases, with correlations strengthening or weakening as the years progress. More formal OLS regressions show negative correlations between migration and consumer spending, employment, and small business revenue. These regressions also show a positive correlation between migration and foot traffic to retail and recreation businesses.

With these insights in mind, I remove the seasonal and trend components of the migration data and use the lagged economic variables to predict the remaining residuals of the migration data. I test several models and find that the optimal model uses all of the economic variables, without county fixed-effects. Using economic variables only presents modest gains in accuracy, as most of the population changes can be modeled with the seasonal variation.

Data Construction¶

This section of code walks through the various datasets and how they are aggregated and merged to form my full dataset.

All of the Opportunities Insights data is relative to a baseline of January 2020. The values for these data are the percent change in the variable from January 2020 levels. In the cases of credit/debit card spending, small business openings, and small business revenues, these values are also seasonally adjusted. Most of the data is either weekly or daily, but the USPS data is only monthly, so I aggregate each of the datasets by taking the mean percent change by month/county.

The USPS data is at the ZIP code by month level. In order to aggregate this data to the county level, I use a ZIP code-to-county crosswalk file compiled by the U.S. Department of Housing and Urban Development.[7] This crosswalk contains the ratio of each ZIP code that lies in each county, by population. I therefore multiply the inflows and outflows by the ratio of each ZIP code that is in each county, then sum these values by county to aggregate the data. Additionally, I use 2019 flows as a baseline in my analysis, so I handle the USPS data for that year separately and then merge the 2019 data as its own column in the eventual dataframe.

I finally merge together the various datasets and calculate the net flows from the USPS and population variables. I have two main netflows variables. The first, 'net_flows_pop', measures the net flows as a percent of the county population. In this case, a value of 1% would mean that in that month, on net, a number of people equivalent to 1% of the county's population in 2019 moved into the county. The second main variable I calculate, 'net_flow_rel2019', performs this same calculation for the 2019 data, then subtracts the monthly 2019 net flows from the 'net_flows_pop' variable. This is meant to account for some seasonality in net flows. For example, people moving from the northeast to Florida during the winter. The 'net_flow_rel2019' variable therefore measures any net flows that are above or below what would be expected based on 2019 data.

Data Visualizations¶

Handling Missing Data¶

First, I examine the missing data in the combined dataset. Given the multiple sources, varied data completeness across these sources, and the timeline of this data, there is likley a substantial amount of missing data. Limiting the analysis to only observations with complete data across all variables could bias the result if counties with missing data are more populous, for example.

To try and understand how data availability may bias my results, I first look at summary statistics for the full dataset and the dataset when conditioning on containing the key economic variables (hereafter: the "subsample"), which tend to have the most missing data. The below tables show these sets of summary statistics, with the first table representing the full dataset.

A key difference that may help explain the other differences between these two tables is in the county population variable. In the subsample, the average county population is over 3x as large as in the full data. This shows that many of the counties being dropped by conditioning on economic variables are smaller counties, possibly where this data is more difficult to collect. The values of the economic variables tend to have lower standard deviations, likely because these variables are relative to January 2020 baselines. In smaller counties with less data, percent differences can be more extreme due to small sample size bias. This is also observed for the net flow variables that are relative to population.

I next explore how the selection of counties may change over time in the subsample. It is possible that over time, the number of counties and/or the composition of counties included in the data changes. This would impact the results, as we couldn't be certain that any changes over time were due to genuine changes in the variable, or if they were due to sample changes. I create the following figures to assess this potential issue:

We see from the first figure that the number of counties in the data is pretty consistent over time. There is a slight drop in observations from 657 to 654, then to 653, but these are minor changes given the sample size. Another concern is that the composition of counties varies over time, though the second figure addresses this concern. Since the number of year-months represented by almost every county is 25, we know they have an observation in each year-month combination. To make this a perfectly balanced dataset, I exclude the 4 counties without complete information over time:

Finally, I look at the geographical distribution of this new list, and find that states along the coast tend to have counties with data, while states in the middle of the country are largely underrepresented. The main purpose of this map is to see where we have complete data, but I also show the preferred net flow variable for June, 2020.

The main dataset I'll use for my analysis is the subsample, given the need for economic variables in order to perform my later regressions. This means that the conclusions I draw from this data are not necessarily generalizable for the entire US population, but will encompass the counties above.

Data relationships¶

I start with a simple representation of data for one county, New York City, over the timeseries of the data. I plot the net migration flows and several economic variables side-by-side to see how closely they follow one another over time. These figures show how NYC experienced a dramatic net outflow during the pandemic, in addition to economic losses across all variables. For most of the economic activity variables, they seem to follow the net flows through 2020 and some of 2021, but the relationship changes when net flows plateau. The economic variables continue to recover, but we don't observe a major return to NYC, which would be represented by positive net flows that equal the negative flows in magnitude. Instead, people just stop leaving NYC in mid-2021. The economic activity variable that most closely resembles this pattern is the GPS data on time spent at retail and recreation businesses, which also starts to plateau at a similar point. This may indicate that net flows are more relevant for predicting economic activty during the pandemic than after, or that we must account for previous net flows when predicting economic activity.

To get a sense of overall trends in the data, I aggregate the monthly data to yearly data. This makes it a bit easier to understand how data in 2020 compares to data in 2021 and 2022 on average. I choose to take the mean of the net flow variables rather than calculating the total net flows by year so that I can compare average monthly changes in flows to average monthly changes in economic variables.

I next consider how county data varies across and within states by looking at ordered boxplots of the yearly data. I plot data for the net flows (relative to population) and credit-debit card spending (rel. January 2020).

These boxplots show that there is a lot of variation of economic variables and migration variables within a state. They also show that the states with the greatest net outflows are not necessarily the states with the greatest drop in credit/debit card spending.

For a more robust analysis of how economic variables may be related to the USPS net flows, I plot the correlations between our primary economic variables and the migration variables by year, at the county-level. This allows me to see which variables have the strongest correlation with migration and whether these correlations change over time.

Time spent at retail/recreation seems to have the most consistently positive correlation with net flows over time. In 2020, there is also a strong positive correlation between small business revenues and net flows, but the regression line becomes much flatter in 2021 and 2022. These changes in correlations over time indicate that certain variables may be more useful predictors in 2020 than in later years.

Given the different effects of COVID across states, which includes the impacts of state policy and COVID prevalence, it is reasonable to think that these correlations may differ by state. To get a sense of this variation, I choose to compare counties in New York and Florida, two large states with different COVID policy and net migration flows. I reproduce the above figure, but just with data from these states, and plot the regression lines corresponding to the data from each state.

It's clear from these figures that the relationships between economic indicators and net flows differ across states, and in some cases even have opposite correlations. In general, the correlations for New York and Florida are more similar in 2021 and 2022 than they are in 2020, likely due to NYC being an extreme outlier in terms of net migration flows. This has several implications for a potential model that would use new migration flows as a predictor of other economic variables. First, the model may need to consider state-level or county-level fixed effects when making its predictions, as the strength and even direction of the relationship between economic indicators and migration varies by state. Second, data from 2020 may be misleading for predicting future trends, given the impact of the initial lockdown. Some version of a time fixed-effect should be used when performing a more robust analysis of the correlations between economic variables and migration flows.

Finally, to get a sense of how net flows vary across the US, I map the yearly total net flows across counties for 2020 and 2021. These maps show similar patterns overall, with a few differences. It appears that 2021 has fewer extreme values than 2020, as observed by lighter shades of green and red. Additionally, zooming in on the counties of major cities shows an exodus from those counties in 2020, with a sustained but less extreme exodus in 2021.